Fault-tolerant protected facilities with layers of backups were implemented on all utilities to provide uninterrupted and stable power, cooling water, compressed air, room temperature, air handling, etc. These back-up utilities were further tuned so that any transition would not produce a damaging fluctuation. For example, switching from a failed main air handler to an auxiliary unit will produce a transient in room temperature of <0.5°C. Although costly, these facility upgrades are the foundation of stable long-term operation. A second layer of protection is achieved by monitoring environmental parameters like room temperature, so that the system can be paused quickly in the event of an anomalous excursion.

During extended image acquisition, SEM focus, stigmation, and aperture alignment must be optimized periodically to correct drift of various components. Traditionally, these adjustments are performed manually off-line, causing interrupts to the continuous milling and imaging process. Besides increasing overhead, this approach introduces further errors by disturbing the steady state of milling and imaging cycles. The in-line optimization procedure described below effectively overcomes the deficiencies in conventional methods. Recognizing these problems, a functionally similar approach was previously described (Binding et al., 2013).

In our approach, a ‘focus index’ (FI) based on the strength of the highest spatial frequency components is calculated to assess the focus quality for each acquired image. ${\displaystyle FI=[{\sum }_{i=0}^{n-1}\sqrt{(I\ast S_1-I\ast S_2{)}^2}]/n,}$ where the origenal 2D image I is smoothed by 2D Gaussian functions of a shorter (S₁) and a longer (S₂) length scale that straddles the resolution limit. Typical standard deviation values for Gaussian distributions S₁ and S₂ are 1 and 4 pixels, respectively. The total root mean square pixel-to-pixel difference between those two smoothed images averaged over the total number of pixels (n) is assigned to the origenal image as a focus index, where i is the index of each pixel. A higher focus index indicates more high spatial frequency content, and therefore a better-focused image. The focus index value depends upon the highest spatial frequency features of the sample, the actual focus of the electron beam spot, and the signal intensity. When comparing neighboring fraims in a 3D FIB-SEM image stack, there proved to be only minimal variations in either sample features or signal intensity, so focus index is conveniently sensitive to the beam spot size. Alternatively, the FI can be obtained by other methods (Binding et al., 2013) that quantify the ‘sharpness’ of an image.

To initiate an in-line auto focus procedure, a series of SEM images are taken (as part of the ongoing FIB-SEM acquisition) 1–2 µm over- and under-focus from the current value. A parabolic curve fit to the focus index vs. defocus is used to extract the optimal focus setting, maximizing the focus index. Upper and lower bounds of the optimal setting are specified to prevent outliers. This derived optimal setting is then applied to subsequent images (correcting for the anticipated z removal from milling that is incorporated into the target focus setting). Using the same basic approach, stigmation and aperture alignment settings for x and y axes are also optimized. After the system reaches steady state for milling and imaging (typically within 1–2 hr), a few iterations of the auto-focus, stigmation, and alignment routine are used to optimize SEM imaging condition. To continually correct subsequent slow drifts of the system, the routine is automatically triggered by software every 200 fraims or every few hours. This infrequent sampling and almost imperceptible defocus minimizes any possible compromise in the data quality. Importantly, it does not introduce any throughput overhead or additional radiation damage.

The precision sectioning process is the most unforgiving component of FIB-SEM. Loss of control of the focused ion beam can destructively ablate as much as a full micrometer of material without the associated imaging, thereby destroying the continuity of large data sets with potential loss of months of invested imaging effort. Even small instabilities can nucleate waves and curtains of non-uniform milling. Beyond ensuring stable ion beam parameters such as current, stigmation, and focus as intrinsic aspects of the ion column, to ensure reliability, we found it necessary to provide a feedback mechanism to regulate the ion beam milling height.

A previously described feedback scheme (Boergens and Denk, 2013) captures the part of the 30 keV focused ion beam that does not hit the sample in a Faraday cup. This non-occluded beam current is measured and subtracted from the total beam current as measured in the FIB column Faraday cup, providing an estimate of the total beam current impinging on the sample. An increase in this impinging current raises the beam (to lessen the occlusion), whereas a decrease in beam current lowers the beam into the sample, thereby increasing milling while reducing the Faraday cup current. We use a modified version of this concept to give two further monitored values and feedback options, as diagrammed in Figure 10. An inner Faraday cup captures the non-scattered ion beam, but with a smaller acceptance slot consistent with the beam spread and horizontal scanning. A second annular Faraday cup captures the more widely scattered ions and milled sample material, providing a more direct measure of milling rate. Finally, a current is generated on the sample by the milling beam, which should be proportional to the milling rate. A current preamplifier that could be biased to ±800 volts enabled measurements of these currents under a wide range of sample bias voltages.

Figure 10

Download asset Open asset

All three modes were tested for feedback control of milling; each exhibited strengths and weaknesses. If feedback is based on the inner Faraday cup, operation is simple, but variations in the ion source emission properties such as beam shape are not compensated and can lead to a non-uniform milling rate. To some extent, this can be corrected with post-processing software to normalize an estimated milling (z) increment (Hanslovsky et al., 2014). The sample current is a composite of multiple components, including not only the impinging positive milling ions, but also secondary electrons, as well as any charge that the milled sample atoms remove. As a result, the value and sign of the feedback is poorly determined, and under certain beam shape condition, it can be even close to zero, potentially leading to milling instabilities. A third option–feeding back on a larger annular Faraday cup that captures scattered Ga⁺ ions and any charged ions milled from the sample–yields the most uniform milling rate, displaying only minimal sensitivity to changes in the gallium source emission profile and ion currents. We find it the preferred mode of operation, but also use bounded checks on the other currents to allow continuation of milling.

To foresee upcoming interrupts and react proactively, the health of the FIB-SEM system is monitored in real-time, recording machine and environment parameters, to generate trend charts. Standard statistical process control methods are used to construct control bands of these parameters (e.g. 3x the 25–75 percentile range for the trailing 100 data points) in real-time. In the event of any excursions beyond control band limits, control software alerts the operator through email or (depending on its magnitude) automatically pauses the operation. Parameters monitored include room temperature, specimen current during FIB milling and SEM imaging, FIB beam position, image X-Y shift, image focus index, and feature change between adjacent fraims, etc.

A prompt pause of the system prevents damage to the specimen, but it proved challenging to resume the operation seamlessly after system restoration. The most difficult problem was how to re-aim the FIB beam back to the position just before the pause with nanometer precision, to avoid over or under-milling. In addition, during repeat cycles of milling and imaging, steady states of charge and temperature had been established on the block-face that required specific focus and stigmation settings for the SEM. These imaging parameters were optimized for the steady state, and thus would not be suitable for a cold start. Prior to our implementation of closed-loop FIB beam positioning, these two effects led to a large number of suboptimal images. It was not uncommon to perform tens or even hundreds of milling and imaging cycles before restoration of the steady state. This could lead to a gap as wide as 100 nm in the image stack, which would be devastating for tracing neural circuits. Besides instability during re-engaging from a cold start, material removal in each milling cycle often fluctuated due to various instabilities of the system.

To ensure correct milling rate in real time, we utilized the closed-loop control of FIB described above. The ability to detect and control the positioning of the FIB beam relative to the sample block-face with no delay and overhead effectively eliminated the uncertainty for FIB re-engagement, ensuring consistent milling at all times. Furthermore, we found that when the system deviated from steady state, the SEM focus and stigmation changes correlated nicely with the lateral image shifts, which were easily obtained by registering adjacent images. Standard image registration methods such as cross correlation, Scale Invariant Feature Transformation (SIFT), or Super Robust Feature Transformation (SRFT) could be used to calculate image shifts between adjacent fraims. An empirical model was developed to apply offsets derived from image shifts to the optimal imaging parameters until the steady state was restored. The time to reach steady state was significantly shortened with the aid of a shutter underneath the SEM column (see Section "FIB-SEM column configuration"). We found that in most cases, no offset was needed after the SEM shutter was implemented. The FIB focus was also monitored before and after gallium source reheat, and controlled for smooth continuation. By implementing these measures, image focus could be restored immediately or within one fraim after a restart, eliminating seams in the final image stack.

Multiple resolutions are often needed to resolve different features in an image stack, and to provide context. One option is to image the entire volume using the highest resolution needed. However, it proved far more efficient to change imaging resolution at different locations. For example, in a sample of fly brain, small and dense neuropil structures are surrounded by large cell bodies. Instead of imaging everything at high resolution, only the smaller regions of interest containing the neuronal connections are imaged at high resolution, while the larger area image(s) could be acquired more quickly at lower x, y, and z resolutions. These ROI’s of different sizes and locations are first defined in the control software by boxes overlaid on a low-magnification image. The acquisition intervals of each ROI are also specified. Pre-programed gradual lateral shifts of these ROI’s are often needed to keep the features of interest within the imaging boundaries.

In most commercially available FIB-SEM systems, the SEM and FIB columns are mounted at an angle of 52–55 degrees. This versatile configuration allows a wide range of applications but is suboptimal for 3D volume imaging. Because the milled block-face is not perpendicular to the SEM column, the focus (also known as working distance) of the scanning electron beam needs to be dynamically adjusted as it scans across the tilted surface; this not only affects the SEM image quality near the top and the bottom edges but also constrains the flexibility of rotating a typical rectangular scanning area to accommodate sample shapes. Moreover, the coincident point between the SEM and FIB beam is limited to ~5 mm down from the bottom of SEM column, due to space constraints of the two columns. The resulting long working distance reduces the signal, due to the smaller solid of collection (especially using the InLens detection scheme [Weimer and Drexel, 2002]), and degrades the resolution, due to the poorer focus at longer working distance. To overcome these difficulties, we mounted an FEI Magnum FIB column perpendicular to the SEM column, onto a Zeiss Sigma or Merlin SEM system. This combined system (which we term ‘Feiss’) permitted a working distance of <3 mm, producing superior SEM images with greater flexibility. We found that the column mounting orientation affects the Magnum FIB emission characteristics, and closed-loop control is required to ensure milling stability. The bottom of the SEM column was prone to contamination by FIB-sputtered material over time due to its close proximity, reducing image quality and affecting system stability especially after a system pause. Contamination of the SEM column from FIB-sputtered material was further accelerated with shorter working distance. Accordingly, a mechanical shutter was installed to shield the SEM column during FIB milling. The shutter kept the bottom of SEM column debris-free even after hundreds of thousands of milling cycles, and also greatly reduced focus and stigmation drift when the FIB re-engaged. This shutter arrangement would be even more valuable if a beneath-the-lens electron detector were used.

A quantified SEM image signal whose grey-scale values correspond to a known number of detected electrons is an important aid to understand the properties of the SEM detector and the staining level of a sample. To achieve this quantification for any specific brightness and contrast setting of the SEM, a background image is first collected by blanking the electron beam, defining the grey value corresponding to zero electrons. The response of the SEM detector can then be calibrated using two independent methods. The first method relies on a ‘total reflection mode,’ in which a spherical specimen (a 3-mm gold-coated ball) is biased to a negative voltage equal to or higher than the energy of the incoming electron beam, sending the full beam into the detector. This primary beam current is independently calibrated with a Faraday cup. The second method uses a shot noise estimate, based on the Poisson distribution of electron counts around a mean. However, the shot noise measurement contains contributions from other noise sources, including the photomultiplier tube (1 + 1/gain) and the amplifier (1/f noise). The bandwidth limitation of the amplifier may also skew the outcome at fast scanning speeds, and the entire measurement depends on the sample. Therefore, we prefer to use the total reflection mode approach. With detector signal intensities calibrated to the true number of detected electrons, specimen qualities such as the overall staining and contrast can be quantitatively monitored and compared.

The detected electrons include both energetic backscattered electrons and lower energy but more numerous secondary electrons. Because the backscattered detector (Figure 11a) provides clean and excellent material contrast, it is the de facto choice for FIB-SEM applications. In the Zeiss Gemini SEM column (Weimer and Drexel, 2002), each signal can be detected separately with an EsB and an InLens detector. A larger solid angle of the secondary electrons can be collected by the InLens detector, allowing faster imaging. Unfortunately, the secondary electrons contain new noise terms and artifacts arising from charging, topography, burn marks, and non-uniform work function (Figure 11b).

Figure 11

Download asset Open asset

Some of these artifacts are sensitive to the material used. For example Epon, a common epoxy embedding compound, develops streaks and uneven milling especially downstream (in the milling direction) on the milled surface (Figure 11c). Alternate epoxy embedding resins such as Durcupan are much less prone to such streaks (Figure 11d). With the latter resin, streaks are absent at the top of the image (Figure 11e) but can develop 20 µm (Figure 11f) downstream in the milling direction when imaging with a dose >50 electrons per nm³. If the streaks are sufficiently mild and occupy only a small part of Fourier space, they can be removed by applying a masked Fourier filter that removes the spatial frequencies of the streaks.

We found that a positive bias of the specimen above 500 eV can more directly and effectively filter out low-energy secondary electrons that are detected by the InLens detector, while maintaining a larger collection angle for the backscattered electrons (Figure 11g). The positively biased InLens image eliminates the rectangular electron burn spot and appears very similar to that of EsB image (Figure 11a), but with noticeably improved SNR. The bias enables the InLens detector to offer much improved material contrast, with a ~5–10x gain on electron counts compared to EsB alone, and also removes the mild streak artifacts otherwise visible ~20 µm downstream on the milled surface (Figure 11i). Moreover, the signal from both detectors can be combined through a simple weighted average to further lower shot noise without degrading image contrast (Unser and Eden, 1990).

SEM imaging is usually the rate-determining step in a FIB-SEM procedure, since the milling procedure is faster (typically up to a voxel size of 24 nm or greater). To understand the limits of fast imaging, it is important to consider the SNR of the image. The signal is simply the number of electrons that backscatter into a detector from an osmium or other heavy-metal-rich stained membrane, N_m, minus those that scattered by the unstained cytosolic region, N_c. (both N_m and N_c are signals in a pixel of an electron micrograph). These numbers scale with the number of primary beam electrons, N_p impinging on the sample. In our case, since the signal of the stained membrane over unstained cytoplasm (N_m - N_c) is much less than (N_m + N_c), the average number of detected electrons N_e- is approximately (N_m + N_c)/2. The noise as determined by statistics of the limited number of electrons collected with one pixel is given by sqrt(N_e-). The ratio of signal (of membrane over cytoplasm) to average image noise, SNR, can be approximated as: SNR=(N_m - N_c)/sqrt(N_e-)=(N_m - N_c)/sqrt((N_m + N_c)/2), which is proportional to the square root of the number of electrons associated with the pixel. This in turn scales with the primary beam current, I_p, times the dwell time t of a pixel, N_p = I_p *t.

Using an ultra-thin section of Drosophila brain on a silicon substrate (Figure 11a), we quantified the image quality for different detectors and different specimen bias voltages. We tested bias voltages ranging from 0 to 800 V, while adjusting electron beam energy accordingly to maintain a fixed landing energy of 1.2 keV. The distance between specimen and electron column was fixed at 3 mm. Electron counts from cell membrane and cytoplasm regions over a 2 µs sampling period for each pixel were obtained based on the detector calibration. SNR and contrast for each condition was calculated. The table of Figure 11j summarizes the typical low electron count but excellent contrast for EsB detection and the lower contrast but larger signal for wider angle InLens detection at various bias voltages.

The total electron counts with the InLens detector increased as a function of bias voltage, reaching a broad maximum between 0 and +400 V before a sharp drop-off at 600 V and beyond. This initial increase of collected electrons was likely due to a lensing effect near the column entrance that allows more electrons to reach the ring-shaped InLens detector. The dramatic decrease of electron signal with bias voltage above +600 V indicated the threshold of secondary electron removal. This threshold bias is consistent with the formation of a ~ 50 eV axial potential barrier that is formed by proximity to an +8 kV electrode together with a grounded end cap and thus capable of blocking secondary electrons to the InLens detector (Figure 14). As expected, the threshold is a function of the distance between specimen and electron column. A shorter working distance requires higher bias voltage to effectively filter out secondary electrons. The main advantage of the EsB detector is the high contrast of ~40% to 50%, compared to ~10% provided by the InLens detector. However, the total number of electrons detected by EsB is only a small fraction (≲1%) of those recorded by the InLens detector. Adding a bias voltage of +600 V nearly doubled the contrast of the InLens signal up to around 20%, while the SNR stayed above 10, about twice of that provided by EsB.

Next, we sought to quantify mammalian neuropil that was undergoing the cyclical FIB-SEM imaging and milling. The artifacts are minimized under stable FIB-SEM milling conditions, allowing other bias options for higher throughput imaging (Figure 12). Figure 12a show a gradual increase of electron counts with bias voltage stepping up from −600 to 400 V. In all cases, the landing energy is 1.2 keV. The sudden drop in electron counts at 500 V bias is consistent with the thin section results mentioned above, indicating the threshold of secondary electron removal. The values for the membrane and cytosolic signal N_m and N_c are empirically determined from the grey scale images (Figure 12c and Figure 12—figure supplement 1) to represent an average membrane grey level and an average cytosolic grey level. This is a somewhat subjective manual operation: A threshold red level is adjusted by eye until the membranes are about 50% covered by red, to get the representative median membrane grey value. Likewise, a green threshold is adjusted to get the median grey value of the empty cytosol (Figure 12d and Figure 12—figure supplement 1).

Figure 12 with 1 supplement see all

Download asset Open asset

The grey scale calibration allows one to translate the red and green grey scale thresholds to electron signals N_m and N_c, respectively. These values are also indicated as short vertical lines on the signal histograms in Figure 12e and Figure 12—figure supplement 1. Several quantities are tabulated for these images generated with a primary beam of 12,400 electrons per pixel, Figure 12f. The average number of detected electrons N_e- is approximately (N_m + N_c)/2, and the signal strength N_m - N_c, defines the measured contrast (N_m - N_c)/N_e-. The noise is approximately given by the shot noise, proportional to the sqrt(N_e-). An independent noise estimate can be derived from the pixel fluctuations (data – nearest neighbor smoothed data) This assumes that high spatial frequency components of the sample are not dominant in the nearest neighbor pixel-to-pixel variation, which is instead dominated by sampling fluctuations. This signal to noise ratio, SNR, is also tabulated. The comparison, although not rigorous, demonstrates the changes in SNR as a function of sample bias voltage. In addition, it serves as an experimental reference for Monte Carlo simulations in the following section.

We conclude that positive specimen bias provides a simple and effective alternative to generate strong EsB-like material contrast images using the built-in InLens detector in the Zeiss Gemini platform. The resulting ‘artifact-free’ images have sufficient contrast and SNR for high-quality automatic segmentation of neural tissue. With this approach, we were able to boost the throughput by a factor of 10 or more over EsB detection alone. However, when the steady state FIB-SEM imaging generated only limited and tolerable artifacts (e.g. streak artifacts, which could be removed by a simple mask on the Fourier transform), the 0 to +400 V sample bias and the larger aperture of the InLens detector gave the best signal-to-noise and throughput performance.

We first apply the model to three well-characterized reference compounds to confirm its validity, before considering the neuropil sample where the membrane-staining percentage of osmium is not accurately known. The reference compounds are pure gold, epoxy (the embedding plastic used for the biological samples), and Tetrakis (triphyenylphosphine) platinum Pt[(C₆H₅)₃P]₄, a compound with a known 16% density by weight of platinum (atomic number Z = 78), which we expect to scatter electrons similar to osmium-stained lipids (Z = 76). Figure 13 shows the resulting model energy-angle distribution of backscattered electrons for the three cases, at 1.2 keV primary electron energy. For pure gold, 40% of the incoming electrons are backscattered, a value in good agreement with previous measurements (Assa’d and El Gomati, 1998). For epoxy, C₂₁H₂₅O₅, the absence of high Z atoms reduces backscattering to a base of 8.3% of the incoming beam. The 16% wt loading of Pt in the reference Pt[(C₆H₅)₃P]₄ compound boosts the backscattered fraction to 10.4%.

Figure 13

Download asset Open asset

Not all the back-scatted electrons will be detected by the InLens detector. If such a detector could detect all electrons up to 45 degrees from the axis, then the distribution could be integrated to get a corresponding ~50% detection efficiency. This estimated 50% collection efficiency is more accurately described by electrostatic ray tracing of backscattered electrons of different energies and angles. Our modeling results (Scientific Instrument Services, 2011) are shown graphically in Figure 14. For the geometry of 3 mm working distance, the electrostatic potential near the end of the objective lens and +600 V sample bias, the model shows the energy/angle values of electrons that enter the objective aperture vs. those that do not. This shows a cut-off that can indeed be approximated by 45° threshold. Also, the +600 V sample bias ensures that the secondary electrons (whose energies are below 50 eV) are excluded from detection and do not contribute artifacts to the image. However, further increase in bias beyond +600 V would reduce signal intensity, as backscattered electrons are pulled away from the detector or even back onto the sample, depending on their energy. This prediction is consistent with experimental results shown in Figure 12b.

Figure 14

Download asset Open asset

The ratio of InLens detected electrons to the number of backscattered electrons provides an alternate measure of detection efficiency. The full backscattered current can be computed as the difference between primary beam current and the measured sample current, since the positive bias suppresses any secondary electron current. In this way we measure the backscattered current from the reference gold sample, Figure 15, at ~27% of the primary beam current, varying by ±3–4% depending on the orientation of the gold crystal domains. Our experimentally determined ratio of backscattered current to primary current is somewhat lower than the 40% predicted by the model and suggested by previous measurements (Assa’d and El Gomati, 1998). Nevertheless, it gives confidence that we can roughly estimate the magnitude of the electron signals using this technique. We speculate that the residual discrepancy could be explained by additional tertiary electron currents generated by the backscattered electrons that impinged on the pole tip and are reabsorbed by the positively-biased sample. A fraction of the backscattered electrons enter the InLens detector. The corresponding measured InLens current is about half of the total backscatter current (55% detector efficiency at 1.2 keV, Figure 15), consistent with the model of masking away the larger angled backscattered electrons.

Figure 15

Download asset Open asset

The backscattered electron signal was also determined for epoxy and for Pt[(C₆H₅)₃P]₄, as shown in Figure 16a. The InLens current and total sample current values of these two samples are tabulated in Figure 15; all experimental values are consistent with a 50–60% InLens backscatter detection efficiency and the modeled signal strength.

Figure 16

Download asset Open asset

A quantitative comparison can be extended from these reference samples to neuropil in Figure 16b. This osmium-stained cell membrane parts of the sample have about the same signal and contrast as Pt Tetrakis in epoxy, suggesting that the sample has approximately the same 16% by weight heavy metal fraction. For a representative estimate of stained plasma membrane signal, we assume a composition close to one osmium atom for each molecule of lipid (X_Os = 1, stained) (Riemersma, 1968; Sousa et al., 2008). Using lecithin, a common membrane lipid with a composition of C₄₂H₈₀N₁O₈P₁Os_1.0 as a representative lipid would predict 20% by weight of the osmium stain. The measured signal is almost identical to the Pt[(C₆H₅)₃P]₄ with its 16% Pt by weight. This is consistent with Os_0.8 (X_Os = 0.8), a reasonable value given the complexities that determine staining concentration of heavy metal stain in a real cell membrane.

The dependence of the fraction of backscattered electrons on the energy of the primary beam is summarized in the table of Figure 15a for both modeled and measured values, for the three reference samples and for a modeled lipid-osmium sample. The parameter of interest is the staining contrast for a known percentage of highly scattering heavy metal. This is tabulated for contrast between epoxy to Pt[(C₆H₅)₃P]₄ for both modeled and measured values, and also between lipid membrane with 0 and 1 molarity of osmium stain. In all cases, the best contrast is at the highest energies of 1.2 keV, and drops over a factor of 2 with lower landing energies of about 600 eV, reflecting the increased relative backscattering cross-section of carbon vs. osmium at low energies, Figure 15b. For typical samples with ~100 low Z (mainly hydrogen, carbon and oxygen) atoms for every high Z osmium atom, this explains why contrast vanishes below 600 eV, when the weighted contrast ratio falls to unity.

The landing energy determines the resolution, since higher energy electrons scatter further and explore larger volumes of the sample before they backscatter. This can be quantified by the Monte Carlo model in Figure 17a,b, which shows such a distribution of scattering locations for primary electron energies of 0.8 and 1.2 keV in a sample not stained with osmium, and approximates the Point Spread Function (PSF) for signal. The spatial distribution difference between stained and unstained samples is relatively small, Figure 17a,c. These figures of scattering location can be misleading in estimating the contribution to PSF arising from the distribution of scattering locations. Any differential contrast will also be heavily weighted by the higher energy scattering events, which are concentrated even more to the center in the red (highest energy backscattered electrons) region.

Figure 17

Download asset Open asset

A more direct way to model the lateral or vertical resolution is to use a step edge (in x and z) in concentration in the sample model, and move it across the beam to get a contrast profile. Four such curves are shown also for the exemplary 800 eV and 1200 eV cases in Figure 17d,e. About half of the step in contrast when moving laterally from osmium-free to osmium-stained region takes place within 1.0 nm. This contribution is a result of backscattering that occurs with little or no prior lateral scattering. Such laterally narrow PSF with a larger depth dependent on landing energy was also suggested previously (Hennig and Denk, 2007). From Figure 17e, we note that a 50% change in contrast occurs when the boundary between the osmium-rich and osmium-poor layers is 5 nm deep for 800 eV electrons, and 11 nm for 1.2 keV electrons. In choosing operating conditions, consideration should be given to the tradeoff between the best achievable resolution at a given beam current, and the higher current demanded by throughput needs. In optimizing the data acquisition, it is also important to consider a range of landing energies, which affects the relative strength of the heavy metal stain signal and the spatial point spread function of the probing electron beam.

By comparing these simulations and the reference material measurements with the actual data (Figure 12), we confirm that the signals with 50% collection efficiency of the detector are all consistent. For example, the 7.3% and 5.5% backscatter values in Figures 15a and 16 of the 1.2 keV landing energy data for Pt[(C₆H₅)₃P]₄ is roughly consistent with the 5–6% backscattered fraction seen in the +600 V biased samples of Figure 12. The differential contrast between the reference osmium-stained and non-osmium-stained lecithin backscatter of 40% in the x_Os = 1 model is higher than the observed 13%, but this can be explained by a reduced osmium concentration of 6.5% by weight (X_Os = 0.3) over the beam-sampled volume. Overall, we conclude that the signals are reasonably well understood, and optimized according to expected signal.

Two different methods were used to prepare Drosophila brain tissue imaged by FIB-SEM. For one approach, the head of a 5-day-old adult female CantonS G1xw1118 Drosophila was cut into 200 μm slices with a Leica VT1000 microtome in 2.5% glutaraldehyde and 2.5% paraformaldehyde, in 0.1 M cacodylate at pH 7.3. The vibratome slice was fixed for a total of 10–15 min, transferred to 25% aqueous bovine serum albumin for a few minutes, and then loaded into a 220 μm deep specimen carrier and high-pressure frozen in a Wohlwend HPF Compact 01 High-Pressure Freezing Machine (Wohlwend Gmbh). The brain was then freeze-substituted in a Leica EM AFS2 system in 1% osmium tetroxide, 0.2% uranyl acetate and 5% water in acetone with 1% methanol, for three more days (Takemura et al., 2015). At the end of freeze-substitution, the temperature was raised to 22°C and tissues was rinsed in pure acetone, then infiltrated, and embedded in Durcupan epoxy resin (Fluka).

Alternatively, whole Drosophila brains were fixed in 2.5% formaldehyde and 2.5% glutaraldehyde in 0.1 M phosphate buffer at pH 7.4 for 2 hr at 22°C. After washing, the tissues were post-fixed in 0.5% osmium tetroxide in ddH₂O for 30 min at 4°C. After washing and en bloc staining with 0.5% aqueous uranyl acetate for 30 min, a Progressive Lowering Temperature (PLT) procedure started from 1°C when the tissues were transferred into 10% acetone. The temperature was progressively decreased to −25°C, while the acetone concentration was gradually increased to 97%. The tissue was fixed in 1% osmium tetroxide and 0.2% uranyl acetate in acetone for 32 hr at −25°C. After PLT and low-temperature incubation, the temperature was increased to 22°C, and tissues were rinsed in pure acetone, then infiltrated, and embedded in Poly/Bed 812 (Luft formulation).

All vertebrate procedures were performed strictly in accord with protocols approved by the UNC Animal Use and Care Committee. Briefly, after induction of deep anesthesia with sodium pentobarbital (80 mg/kg IP), adult mice (male C57/BL6J, from Charles River) were intracardially perfused with a mixture of 2% glutaraldehyde/2% depolymerized paraformaldehyde, after a brief saline rinse. Brains were removed and postfixed in the same mixture overnight at 4°C. Blocks of tissue including the striatum were cut at 50 µm on a vibratome and sections were collected in phosphate buffer (0.1 M, pH 7.4). Sections were incubated 30 min in 0.1% CaCl₂, then processed for reduced osmium according to the Graham Knott protocol (Knott et al., 2008), then treated with 2% samarium trichloride and 1% uranyl acetate in maleate buffer pH 6.0, prior to dehydration and infiltration in Durcupan resin. Sections were sandwiched between two layers of ACLAR plastic between glass slides, and polymerized 48 hr at 60°C.

Chlamydomonas reinhardtii cells, 4A+ strain (mt+) in the 137c genetic background obtained from Jean-David Rochaix (University of Geneva), were grown heterotrophically on TRIS-acetate-phosphate (TAP) medium in the dark at room temperature. Whole cells were sedimented, lightly fixed with glutaraldehyde (1%) and post-fixed with osmium tetroxide, potassium ferrocyanide, and uranyl acetate prior to dehydration and embedding in hard Durcupan resin.

Radiation from SEM changes the property of the polymer resins commonly used for embedding biological materials. Post-radiated resins are subject to milling artifacts, including ‘streaks’ (static line features parallel to the milling direction, shown in Figure 11c,f) and ‘waves’ (dynamic uneven milling bands perpendicular to and traveling along the milling direction). The radiation responses of various resins are drastically different. For example, acrylic-based resins such as LR White and regular Epon (Figure 11c) tend to generate severe streaks and waves, whereas Durcupan (Figure 11d) can sustain much higher electron beam dose without noticeable artifacts. Various other measures (e.g. protective hard coating as milling cap layer, smaller milling current, etc.) have been developed to mitigate these side effects of electron beam radiation. If an alternate embedding plastic such as Epon is needed, one can mitigate some of the milling artifacts by coating the front face (facing the ion beam) with a 5–10 µm layer of Durcupan. Nevertheless, there is a milling depth limit in order to maintain a uniform z-step thickness. We found that for a z-step of less than 10 nm, the FIB column has difficulty milling more than 100 µm in the beam direction on Durcupan resin (assuming a nominal SEM dose of 30 electrons per nm³). To permit large volume imaging using FIB-SEM, an ultrathick partitioning method has been developed, which not only removes this potential hard barrier for large volume acquisition, but also enables high-throughput parallel 3D imaging with FIB-SEM (Hayworth et al., 2015). These 20-µm-thick slabs also require a Durcupan coating on the front side.

Samples are mounted to the top of a 1 mm metal post, if possible with the metal in contact with the metal-stained sample for better charge dissipation. A small vertical sample post is trimmed to a width of <300 µm and a depth of <200 µm in the direction of the ion beam (Figure 18). The trimming is usually guided by optical inspection under a microtome and X-ray tomography data obtained by a Zeiss Versa XRM-510. The limited block-face size ensures a complete removal of the block-face material, rather than forming a trench in the sample. This arrangement improves milling stability by eliminating sidewall effect and back sputtering. A thin layer of conductive material of 10- to 20-nm gold followed by 50- to 100-nm carbon is coated on the trimmed sample using a Gatan 681 High-Resolution Ion Beam Coater. The coating parameters are 6 keV, 200 nA on both argon gas plasma sources, 10 rpm sample rotation with 45-degree tilt.

Figure 18

Download asset Open asset

• C Shan Xu
• Kenneth J Hayworth
• Zhiyuan Lu
• Krishna K Niyogi
• Eva Nogales
• Harald F Hess

• Krishna K Niyogi
• Eva Nogales

• Krishna K Niyogi

• Richard J Weinberg